liamsi/a2c-PandaReachJointsDense-v1

🦾 PandaReachJointsDense-v3 with A2C

This agent was trained using Advantage Actor-Critic (A2C) from Stable-Baselines3, designed to solve the robotic manipulation task PandaReachJointsDense-v3. Training was logged using Weights & Biases.

🌎 Environment

• Gym ID: PandaReachJointsDense-v3
• Task: Robotic manipulation (reach joints) with dense rewards

🧠 Model Configuration

• Algorithm: Advantage Actor-Critic (A2C)
• Policy Network: Multi-Input Policy
• Learning Rate: 0.0003
• Gamma (Discount factor): 0.99
• Number of steps per update (n_steps): 5
• Entropy Coefficient (ent_coef): 0
• Max Gradient Norm: 0.5
• Total Training Timesteps: 500000

📊 Evaluation Details

Evaluation was performed regularly during training (every 1000 timesteps), and the best-performing models were saved according to the evaluation metrics.

• Evaluation Frequency: Every 1000 timesteps
• Evaluation Mode: Deterministic policy
• Evaluation Environment Wrapped: Monitor

📈 Logged Metrics on Weights & Biases

• Policy Loss
• Value Loss
• Episode Reward (dense reward)
• Episode Length
• Entropy (to measure policy exploration)

🔗 Useful Links

• Weights & Biases Project

Enjoy exploring and using the trained agent! 🦾✨